2026年3月3日星期二 ➢ 伊朗战争刺激美通胀担忧；英伟达 H200芯片或对华限购。A股午后冲高后 维持强势 。 上 证 指 数 收 涨 0.47% 报 4182.59点，深证成指跌0.2%，创业板 指跌0.49%。油气板块多行业领涨大盘。 避险情绪引发港股大挫。恒生指数收跌 2.14%报26059.85点，恒生科技指数跌 2.89%，恒生中国企业指数跌1.78%。油 气行业大爆发，贵金属板块涨幅居前。 大市成交额大放量至3576.787亿港元。 外盘方面，欧洲三大股指全线收跌，美 国三大股指收盘涨跌不一，道指跌 0.15%报48904.78点，标普500指数涨 0.04%报6881.62点，纳指涨0.36%。美国 称对伊朗最严厉打击尚未到来，否认会 陷入持久战。据悉美国考虑出台英伟达 H200芯片限购令，每个中国买家购买 上限为7.5万枚。 | 26059.85 | -2.14 | 1.67 | | --- | --- | --- | | 8701.91 | -1.78 | -2.38 | | 4989.37 | -2.89 | -9.55 | | 120.04 | 1.58 | -2.77 | | 41 ...

环球市场动态 欧 盟 经 济 安 全 战 略 更 新 股 票 受伊朗局势影响，全球股市整体下 跌。A 股周一低开高走，表现坚挺， 油气股大幅拉升；港股大跌，科技 股表现疲软；欧洲股市全线回调， 国防军工板块逆市上涨；美股表现 分化，科技股走强带动标指基本持 平，AI 相关个股全线回暖。 外 汇 / 商 品 中東局勢升溫，周一国际原油价格 飙升 6%；投资者担忧能源价格飙 升可能会刺激通胀，黄金收窄涨幅 至 1%。 固 定 收 益 美国国债全线走弱，主要受到地缘 政治与通胀预期升温的双重影响。 投资者担心中东局势可能持续发 酵，美国政府或需要增加融资以支 撑潜在的延长军事行动。此外，美 国宏观数据亦偏强，加大了美债的 下行压力。 产品及投资方案部 注：bp/bps=基点；pt/pts=百分点 中信证券财富管理 (香港) 免责声明请参考封底 2026 年 3 月 3 日 ▪ 内外形势重大变化之下，欧盟于 2025 年 12 月发布了新版的《加强欧盟经济安全》框架，在 "去风险" 姿态上 更为主动。第一，政策执行尺度方面，受限于政策约束力较弱及成员国间存在利益诉求差异，政策实际执行上 难以同美国一样整体极端推进。对 ...

【导读】人民币继续升值，A股震荡 今天的A股市场整体偏震荡，但持有恒生科技指数的小伙伴们估计体感就稍微差一些，希望你们坚强挺 住！ 一起看看发生了什么事情。 2月26日，市场全天探底回升，三大指数涨跌互现。截至收盘，沪指跌0.01%，深成指涨0.19%，创业板 指跌0.29%，科创综指表现亮眼，涨超1%。 市场共2485只个股上涨，其中78只个股涨停；2870只个股下跌。 | 880005 涨跌家数 | | | | | --- | --- | --- | --- | | 日中 | 涨停 | | 78 | | 涨幅 | > 7% | | 152 | | 涨幅 | 5-7% | 1 | 133 | | 涨幅 | 3-5% | | 313 | | 涨幅 | 0-3% | | 1887 | | 跌幅 | 0-3% | | 2646 | | 跌幅 | 3-5% | | 179 | | 跌幅 | 5-7% | | 32 | | 跌幅 | > 7% | | 13 | | 其中 跌停 | | | 5 | | 上涨家数 | | | 2485 | | 下跌家数 | | | 2870 | | | | | 130 | | 平盘停牌 ...

3. 美国总统特朗普在国会发表国情咨文演讲，用时108分钟，创历史之最；特朗普称，已从全球范围内 获得18万亿美元的资金承诺，美关税将依据其他法律条款继续维持有效。 (格隆汇) 美股时段值得关注的事件： 1. 美国三大股指期货小幅上涨，纳指期货涨0.39%，标普500指数期货、道指期货涨0.31%。 2. 欧股主要指数集体上涨，德国DAX指数涨0.28%，英国富时100指数涨0.93%，法国CAC指数涨 0.38%，欧洲斯托克50指数涨0.69%。 4. 穆迪首席经济学家拉响警报：市场与经济脱钩，一场大规模、全资产抛售将至。 5. 美国联邦住房金融局：房利美、房地美"非常有可能"今年上市，估值5000亿至7000亿美元。 6. 据报美英两国讨论民用核技术的合作，商讨联合举办聚变技术峰会。 7. 美国商务部：英伟达H200芯片对华出口因流程中设置的防护措施仍处于停滞状态。 8. 谷歌重押"铁空气电池"，宣称能提供长达100小时的稳定电力。 9. 派拉蒙收购报价加码至每股31美元，华纳兄弟：或优于奈飞现有方案。 10. 美国铝业寻求将其关闭或缩减规模的10个站点出售予数据中心产业。 11. Unity软件据报考虑 ...

直接把大模型装在芯片面上。。。 这样什么内存带宽限制之类都不存在了。。。 而且性能非常夸张： 在这块专用芯片上跑Llama 3.1，速度能达到17000 token/秒。。。 英伟达最强的H200大概230 token/秒。。。 英伟达超级大芯片B200也就2000 token/秒。。。 来源：贩财局 这个春节，有个叫Taalas的芯片新公司横空出世，它干了一件让英伟达震惊的事情： 这个差距不是一点半点，是直接降维打击。。 想体验可以到这里：chatjimmy.ai 我试了一下，感受到快、非常快、恐怖的快。。。不亚于第一次使用ChatGPT的震撼 按照现在发展速度，很快实现22,000 tokens/s，意味着AI的反应速度已经超过了人类神经的传导速度。 我总结它的优点： 每次升级都得换硬件。。。。 打个比方，传统GPU像是任天堂Switch游戏机，卡槽一插，想玩《塞尔达》玩《塞尔达》，想玩《马里 奥》玩《马里奥》，换游戏只需要换卡带。 Taalas相当于直接把《塞尔达》这部游戏烧死在主板上。想玩《马里奥》？没法下载，也没法插卡，只 能把整台机器扔掉，买一台出厂就烧好《马里奥》的新机器。 不管怎么样，Taal ...

特朗普重返白宫后，再度挥舞关税大棒，试图以极限施压迫使中国让步，美国前财政部长顾问给他泼了 一盆冷水。奥巴马时期担任财政部长顾问的史蒂文·拉特纳，在结束一周的访华行程后，于美国主流媒 体《纽约时报》发表署名文章《我刚从中国回来，我们没有赢》，直言不讳地指出"中国制造业实力强 大，不可能通过加征关税打败中国"。 这番打破美国政客固有论调的表态，是美国精英阶层对中国地位的正视与承认，还是暗藏私心、为美国 后续对华策略"铺路"的话术？ 史蒂文·拉特纳 关税战终成"纸老虎" 特朗普自开启第二任期，对华经贸政策迅速回归"极限施压"轨道，以近乎"拍脑门拉表格"的草率方式， 炮制出所谓"对等关税"，甚至荒唐到对只有企鹅栖息、无人类居住的岛屿商品加征关税。 作为美国眼中"最主要的竞争对手"，中国自然成为这场关税闹剧的重点针对对象。美国率先对中国商品 加征34%关税，还蛮横威胁中国不得反抗，否则将进一步加码制裁。但中国从不惯着霸权行径，美国加 征多少，中国便同步反制多少，坚定将"对等"二字贯彻到底，绝不妥协退让。 拉锯之下，中美双方关税一度攀升至145%，近乎宣告双边正常经贸往来陷入停滞，而美国国内随即陷 入混乱：民众疯狂抢购 ...

Core Insights - The article discusses the increasing demand for AI Data Centers (AIDC) driven by the exponential growth in computing power requirements due to generative AI advancements and supportive government policies like "East Data West Computing" [5][32] - Major tech companies are ramping up investments in AI infrastructure, with ByteDance planning to increase its capital expenditure to approximately 160 billion RMB in 2026, while Alibaba aims to invest over 380 billion RMB in technology R&D and infrastructure over the next three years [7][34] - The article highlights the penetration of AIDC into traditional industries, evidenced by significant procurement projects such as China Mobile's purchase of 7,499 AI servers for 2025-2026 [8][35] AIDC Types and Characteristics - AIDC is categorized into three types: General Data Centers, Intelligent Computing Data Centers (AIDC), and Supercomputing Data Centers, each serving different computational needs [4][30] - General Data Centers focus on traditional data storage and management using CPU servers, while AIDC leverages AI chips like GPUs for large-scale model training, and Supercomputing Data Centers support advanced scientific research [4][30] Five-Layer Cake Theory - NVIDIA's CEO proposed a "Five-Layer Cake" structure for AI infrastructure, which includes Energy Layer, Chip and Computing Layer, Infrastructure Layer, AI Model Layer, and Application Layer [10][37] - The Energy Layer is crucial for providing stable power to AIDC, while the Chip and Computing Layer focuses on high-performance hardware [11][39] - The Infrastructure Layer integrates energy and chip resources to deliver intelligent computing services, and the AI Model Layer is essential for developing models that drive AI applications [13][41] Industry Ecosystem and Opportunities - The AIDC industry's growth is a result of the synergy between computing power demand and technological advancements, benefiting various sectors [18][45] - The transition to high-voltage and direct current power systems is becoming mainstream, with NVIDIA introducing an 800V DC power architecture to meet the power demands of next-gen AI facilities [19][46] - Liquid cooling systems are gaining traction due to the high power consumption of AI servers, leading to increased market demand for cooling technologies [20][47] Domestic AI Chip Market - The domestic AI chip market is diversifying, with multiple brands achieving significant sales volumes, indicating a shift from technology development to large-scale delivery [25][52] - The price range for domestic AI inference chips is between 30,000 to 200,000 RMB, with a notable increase in production expected as manufacturing capacity improves [25][52] Conclusion - The article emphasizes that while China has advantages in energy resources and computing infrastructure, breakthroughs in high-end chip development and core technology innovation are still needed [26][53] - The ultimate winners in the AI industry will be those who can integrate full-stack technologies and foster collaborative industrial advancements [26][53]

Core Viewpoint - ByteDance is developing an AI chip and negotiating with Samsung for chip manufacturing to secure advanced processor supply [1][2] Group 1: Chip Development and Production Plans - ByteDance plans to obtain chip samples by the end of March and aims to produce at least 100,000 chips this year, with a potential increase to 350,000 chips [1] - The chip is designed specifically for AI inference tasks, indicating a strategic focus on enhancing AI capabilities [1] - The project, codenamed SeedChip, is part of ByteDance's broader commitment to AI research and development [2] Group 2: Competitive Landscape - Competitors like Alibaba and Baidu have already made significant advancements in chip development, with Alibaba releasing the Zhenwu chip and Baidu selling its Kunlun chip [2] - ByteDance's investment in AI-related procurement is projected to exceed 160 billion RMB this year, with over half allocated for purchasing NVIDIA chips and advancing its self-developed chip project [2] Group 3: Industry Context - The global AI infrastructure is experiencing a supply shortage for storage chips, making ByteDance's collaboration with Samsung particularly attractive [1] - Major tech companies, including Google and Microsoft, have also pursued self-developed AI chips to reduce reliance on external suppliers, with NVIDIA currently dominating the advanced chip supply for AI development [1]

Core Viewpoint - ByteDance is developing an AI chip and negotiating with Samsung Electronics for manufacturing, aiming to produce at least 100,000 chips designed for AI inference tasks by the end of March, with plans to increase production to 350,000 units later this year [1] Group 1: AI Chip Development - ByteDance plans to receive chip samples by the end of March and aims to produce at least 100,000 AI chips this year, with a gradual increase to 350,000 units [1] - The negotiations with Samsung also include securing a supply of storage chips, which are currently in high demand due to the global AI infrastructure boom [1] Group 2: Investment in AI - ByteDance intends to invest over 160 billion RMB in AI-related initiatives this year, with more than half allocated for purchasing NVIDIA H200 chips and advancing its self-developed chip program [1] - The company has initiated a project named "SeedChip" as part of its broader efforts to direct resources towards AI development, believing that this technology will transform its business portfolio, which includes short videos, e-commerce, and enterprise cloud services [1]

Group 1 - The current Trump administration's "mercantilist" approach is evident in its chip tariff strategy, which is more about negotiating with chip manufacturers than a definitive goal [1] - Reports indicate that the U.S. plans to exempt major tech companies like Amazon, Google, and Microsoft from upcoming chip tariffs due to their investments in AI data centers, with exemptions linked to TSMC's investment commitments [1][4] - The U.S. government has announced a 25% tariff on certain imported semiconductors and related equipment, citing national security threats, while noting that only 10% of the chips needed are produced domestically [3] Group 2 - The new plan allows TSMC to allocate exemption quotas for its U.S. clients based on their investment levels in the U.S., aiming to encourage more domestic chip manufacturing [4][5] - TSMC has committed to investing $165 billion in expanding its U.S. production capacity, with the U.S. government closely monitoring the integrity of tariff and rebate measures [5] - Major U.S. tech companies are projected to invest $700 billion in AI infrastructure by 2026, nearly double last year's total, with significant portions allocated for data centers and advanced chips [6] Group 3 - The AI investment surge is reshaping the economic landscape, leading to increased demand for memory chips and computing components, which is causing shortages in parts needed for traditional consumer electronics [6] - European investments in AI, such as the €20 billion announced by France, are significantly lower than the investments by U.S. tech giants, highlighting challenges in financing AI development in Europe [7] - The reallocation of labor and materials in the U.S. to support large tech projects is driving up construction costs and exacerbating housing crises in various cities [7]